High-Resolution Wetland Indentifcation through Machine Learning

Applications to Forested Wetlands in the Hiawatha National Forest

wetlands-hiawatha-1.jpg

A small, ephermeral wetland in a Lake Michigan shoreline 'dune and swale' complex in the Hiawatha National Forest

Why do we need good wetland maps?

Accurate, high-resolution wetlands maps make it easier for interdisciplinary land management teams to communicate about wetland conservation, and therefore better enact sustainable land managment plans.

In U.S. national forests like the Hiawatha, as well as on other public and privately-owned forested properties, the land is managed for human uses including recreation and timber harvesting. However, forests are generally healthier (and therefore harvests and recreating are better!) if these areas are sustainably management to support natural ecosystems, like wetlands.

But what are wetlands and why is it important to minimize our impacton them?

According to the EPA,wetlands are areas where water covers the soil, or is present either at or near the surface of the soil all year or for varying periods of time during the year. Wetlands are important environments for many reasons. Two often-sited reasons are that they support high concentrations of animals and serve as nurseries for many species and that they provide a range of ecosystem services that benefit humanity, including water filtration, storm protection, flood control and recreation (World Wildlife Fund). If wetlands are disturbed, filled in, or otherwise destroyed they can no longer serve these functions. The effects of such distrubances also have a global impact, since wetlands are important to the global migration of animals, such as birds and act as sinks for atmospheric carbon (Nahlik and Fennessy, 2016).

Wetlands are fragile and susceptible to human alteration. Because it is difficult to build or farm on wetlands, they have often been treated as an inconvience by people, which has lead to the alteration of many wetlands by people. Other human impacts on wetlands are wholly inadvertant. Some examples of human impacts on wetlands include alterations to flow regime, changes in weater chemistry, habitat alterations, and the introduction of new species (Queensland Department of Agriculture and Fisheries). Specifically in Michigan, ~40% of wetlands have been destroyed since the early 1800's (FLOW).

Simply put,

**better maps = better land managment = healthier wetlands = healthier forests**

Why do you want to make "better" maps? What's wrong with the maps currently being used?

Some wetlands map are not very high resolution - which allows for wetlands smaller than ~60 m (200 feet) in diameter to be left off maps.

Like any type of map, the maps of wetlands used for land management purposes in the U.S. vary in quality, accuracy, and amount of detail provided. This variety usually depends on either the original purpose of the map - which would determine its extent and scale (resolution) - and the type of data that were availible when the map was made.

When I worked at the Hiawatha, the map of wetlands used during forest planning was the U.S. Fish and Wildlife Service's National Wetlands Inventory (NWI). This is an excellent, national, polygon-based dataset that provides detailed information on the abundance, characteristics, and distribution of U.S. wetlands. NWI data can be accessed through the Wetlands Mapper tool. The completeness and accuracy of the NWI varies by region (Chignell et al., 2018).

While the NWI dataset was originally created in the 1970's and 1980's using high-altitude (1:130,000, 1:80,000, and 1:62,500 scale) photographs; 1 mm on these images would equal 120 m, 80 m, and 62.5 m respectively. While this original mapping was a monumental and important effort, this ~120 - 62 m spatial resolution did not always allow for mappers to capture smaller-scale wetlands. These wetlands datasets tend towards, errors of omission rather than commission, particularly in remote and forested areas, like the Hiawatha NF (Chignell et al., 2018).

Fortunately, the NWI database is constantly being updated, using increasingly detailed dataset, with new versions released in May and October of every year!

Unfortunately, which areas of the U.S. are updated is a piecemeal effort, generally based on local or state funding initiatives. Many of these area aren't updated more frequently because that process is time-intensive, requires expert skill and knowledge, and is therefore expensive.

Proposed Solution

The purpose of this project is to try to accelerate the high-resolution (1 m- scale) classification of wetlands using machine learning with remotely sensed datasets.

Study Area: Hiawatha National Forest

The Hiawatha National Forest is located in the eastern Upper Peninsula of Michigan, USA (see map below). The Hiawatha is devided into two "zones" - the East Zone and West Zone. Some watersheds within the Hiawatha NF have updated, 1m-resolution NWI datasets (like the Brevort-Millequonquins watershed - see map), while others do not (like the Carp River watershed).

The goal of this study is to make a 1m classified wetlands raster of the Carp River Watershed, based on machine learning in adjacent watershed.

Map of the Hiawatha National Forest and the watersheds proposed to be used in this study. The thick black outlines indicate the boundaries of the Hiawatha National Forest. The blue area shows the Carp River Watershed, and the darker blue areas are the NWI wetland polygons within that watershed. The orange areas shows the Brevort-Millequonquins Watershed, and the darker orange areas are the NWI polygons for that watershed. You can zoom in and click on all the polygons to see what their wetland type is!
These histograms show frequency of (Area/Perimeter) values for NWI polygons in the Brevort-Millequonquins and Carp River watersheds. The smaller the (Area/Perimeter) value, the more complex the polygon shape is. This simple parameter demonstrates that the older NWI data of the Carp River Watershed is less detailed than the more recently mapped data of the Brevort-Millequonquins.

Work Flow

wetland-classification-flow-chart.png

Proposed project work flow. Boxes are color-coded as follows: Green = Done/Almost Done, Red = Might get to this semester, Grey = Probably won't get to.

Data Types and Sources

Likely model input datatypes:

Other Possible Inputs:

There are many other nation-wide datasets that can be used to indentify wetlands. These datasets generally have a coarser resolutions, and would have to be down-sampled to be applied to this study.